Sponsored: by Google Cloud | Architecting the End-to-End Stack: Scaling AI Without Scaling Cost

Summary

Google Cloud and Databricks integrate through bidirectional Delta Lake and Iceberg interoperability, with Google's Axion C4A processor delivering 40% better price-performance for Databricks workloads.
InMobi scales AI across 320 billion ad queries per day using up to 10,000 peak Spark nodes on Google Cloud and Databricks, maintaining tight cost discipline to sustain efficiency at that volume.
After adopting a unified agentic AI pipeline on Google Cloud and Databricks, InMobi achieved 80% fewer duplicate jobs and a 20% reduction in compute costs.

Chapters

FAQs

What makes the Google Cloud and Databricks partnership uniquely valuable?

The partnership combines Databricks' data and AI capabilities with Google Cloud's infrastructure—including the Axion C4A processor, which delivers 40% better price-performance for data workloads—plus deep integrations across Vertex AI, Knowledge Catalog, and Unity Catalog federation. Customers benefit from a unified stack that avoids vendor lock-in through an open-source foundation with Delta Lake and Iceberg interoperability.

How does InMobi scale AI workloads with Google Cloud and Databricks?

InMobi processes 320 billion ad queries per day using a peak of 10,000 Spark nodes on Google Cloud and Databricks, with a disciplined architecture focused on cost efficiency at that scale. The company chose this stack over legacy infrastructure specifically for its ability to handle extreme data volumes while keeping compute costs manageable.

What results did InMobi achieve after adopting a unified agentic AI pipeline?

InMobi achieved 80% fewer duplicate jobs and a 20% reduction in compute costs after consolidating on an agentic AI pipeline built on Google Cloud and Databricks. The company also accelerated MLOps cycles, moving models from experimentation to production in days rather than weeks.

How does Unity Catalog federation support governance across Google Cloud and Databricks?

Unity Catalog federation provides a single governance layer that allows teams to query sensitive records across platforms without requiring data movement. This video presents it as a 'single pane of glass' for governance, enabling organizations to enforce consistent access controls across multi-cloud environments.

Full transcript

[00:08] Good afternoon everyone. It's 1:50. I think it's about time. Folks might trickle in, but I'll get started. Um, first off, thank you for coming. Um, it's great to be in front of the data bricks community and talking about
[00:20] Google cloud and data bricks. Uh, today we are going to talk about some of the toughest challenges we see in the industry. How do you scale AI as well as keeping your costs and managing your
[00:32] cost effectively? And in some sense um we'll give you a quick view of how not just like our customers in moi in
[00:44] particular has traversed this journey um today. So that's going to be a key part of the presentation today. Uh I am Naven Punjabi. I lead our data cloud and AI partnerships at Google cloud. I've been
[00:57] working with data bricks for quite a few years and standing on the shoulders of my team. Um we'll be joined by Mohit Sakana who is the group CTO at InMobi who's going to really dig deeper into
[01:10] the journey that they have taken with Google cloud and data bricks together to scale their AI workload. So that's going to be something I'm the opening act if you will. Um before we dive deep into the journey
[01:23] of InMobi, I wanted to share context about the momentum that we are seeing with data bricks on Google cloud and what are the key undercurrens behind it. Uh we announced our partnership with
[01:35] data bricks in 2021. We made it available on Google cloud and since then we have seen a lot of momentum 2,000 plus customers using data bricks on Google cloud. now available across
[01:48] multiple regions 15 plus including the serverless products uh whether it's workflows whether it's data brick SQL all of the capabilities are available on Google cloud for you to be able to leverage and one of the things that we
[02:02] have seen as we have engaged with customers is you think about why they are bringing leveraging Google cloud and data bricks uniquely together they think about it not just from a price performance perspective that hey is it
[02:15] cheaper prefer to run it on Google cloud but they also think about the value it drives in terms of the key integrations and the unique capabilities data bricks and uh Google cloud bring together so
[02:27] that's one of the key underpinnings Thomas Kuran and Ali goody have been our product sponsors and they have leaned in heavily in terms of guiding the path forward for this partnership in terms of the key pillars when we
[02:41] think about data bricks and Google cloud what is unique why is Is it different? It starts at the infrastructure layer to begin with. That's the core underpinning. How do we bring together
[02:53] better price performance for data bricks workloads on Google cloud? It starts with lot of the work that we are doing with custom silicon on our end on axion C4 instances. That's a key driver. We'll
[03:07] spend some more time on in terms of what it means with respect to chip, whether it means bandwidth, what it means in terms of disk storage and there is a separate session that we actually did in
[03:19] the morning to cover just on that topic. The second key pillar is around openness. Data bricks is a key driver within the open source ecosystem and they've been closely collaborating with Google across the entire journey.
[03:32] Started with Spark also on Kubernetes. So data bricks was available even using our GKE instances at that point in time in terms of and then how that has taken forward with ML flow with delta we're
[03:45] closely collaborating within that space. The third key element is Google also has a lot of key capabilities in the analytics realm and what we have done really well together is worked with data bricks to drive some of the core
[03:58] integrations with our platform. So on the data side, how do you bring together data effectively into data bricks on Google? Whether it's using pubsub data flow, if you're doing insights on top of data bricks data and if you're using
[04:12] looker, how can you do that effectively? And we have continued to move that journey forward as more and more products come out both on the data brick side as well as on the Google side. The third key element which I think is
[04:24] critical for us is AI is happening very quickly and there's a lot of investment Google has been making on the AI stack itself. So if customers are building models on agent bricks and how do you
[04:38] serve that effectively within the vortex stack is a key component. We're also bringing together 200 plus models available through our vortex platform. How do you serve that effectively with data bricks so that customers get
[04:51] advantage of both of them? That's the fourth pillar in terms of enterprise AI. This one is a critical element of the open stack that we talked about. We
[05:03] announced this a few months ago at Google Cloud Next around birectional integration at based on the iceberg standard at u.
[05:15] So what it means is at the storage layer if you're if you're using Google cloud you have all your data that you might have within different systems available it can be shared with data bricks in an
[05:27] iceberg format. So there is interchangeably interchangeability there. We have seen customers who have data in multiple application systems. They use Google cloud as the core storage layer bring all of that data together into iceberg and interoperate
[05:40] with data bricks. We have also seen customers who do who who are doing the opposite where bring they're bringing the data together into the apps ecosystem within Delta Lake and then sharing it for specific workloads on the
[05:53] Google side. So I think this birectional interoperability at the storage layer is now the core fulcrum of how things go and it's helping really customers remove the data silos. One of the things that we announced at next was how we are also
[06:07] bringing this interoperability crosscloud. So you might have some of their data assets in iceberg format across AWS across Azure and we now have capabilities to bring all of that data
[06:19] together in some ways making it seamless for customers to say I need a data platform. How do I bring my data into driving insights together? So this is one there will be more investments that we'll continue to do in this area
[06:31] because how do you bring together all of this capability not just at the storage layer but at the catalog layer and then how do you bring that unified context for AI that's what we are solving for so
[06:43] you'll see some good investments from us in this in this domain action this is the price performance aspect of what we are doing there we data bricks now when it runs on Google
[06:55] cloud heavily leverages the action platform from we announced C4A the C4A instances which is uh 40% much more efficient as compared to our second generation. It is optimized for spark
[07:08] execution. It is optimized for data inensive workload. So hence you will see higher price performance and when we've had customers using this capability on Google cloud many of them have run into 40% more efficiencies than ever before.
[07:21] So this is a critical element. It also comes together not just at the silicon layer. Uh think about when you use this capability along with hyperdisk that we bring together which has high higher IO
[07:34] capability. It also adds more and provides lower TCO as well as higher value to our customers and then of course bundled together with enhanced networking the private IP capability that Google clouds brings together. uh
[07:48] whenever you're looking at thinking about data bricks workloads running more efficiently on cloud you'll see data bricks uh on Google cloud front and center but now thinking about AI right so this
[08:01] is great you're able to bring together the best of Google as well as the best of data bricks but how are we helping together solve challenges around AI this is a key component that we run into when we talk to customers they talk a lot
[08:14] about taking workloads that they are running in P and experimentation. How do you take it into production scenarios but when they do that they're also running into investments that they've
[08:26] made in their legacy infrastructure. How do you bring it together into cloud and do governance around it? So the three challenges you'll see is like taking AI workloads from production to scale. Second one you'll run into is governance
[08:39] and how do you provide a modernization strategy that works from where you are to where you need to go. So in some sense Google um has been doing this internally for quite some time within
[08:52] our own capabilities and it's not just working with customers they want customers want partners who are gone through this journey and we are doing the same in Google cloud some key data points for you to look at is 75% of the
[09:05] code that is now being generated within Google is also AI generated in many ways or AI supported and the other key element is if you think about threat and security ops teams are you are using
[09:17] agents to actually look at the threats that are coming in and are reducing this u turnaround time for Google to be able to serve those requests by 90%. So that's a lot more efficiency that's
[09:29] coming in. This is Google doing before we roll it out to our customers. How are we scaling workloads effectively within Google? And now this over time we're making this more and more available to
[09:41] our customers. And the customers are loving it. We have 8.5 million plus developers building on our stack. Nine of the 10 AI labs are using Google Cloud for their training and inference
[09:53] purposes. So that really shows the capability of the scale that Google is able to bring together on the AI infrastructure side to help you succeed and in some ways before I pass it to Inmoi because that's the journey that I
[10:05] think everybody is interested more in terms of how do you take this capability forward and make it available to solve for your customer needs. The thing that I want to leave you with is Google is
[10:17] the only uh cloud provider that has full endtoend stack for AI starting from the silicon start next to the AI research team that we have who's building the models together. How do you bring the
[10:30] data cloud open data cloud aspect along with data bricks in addition to the acquisitions that we did with whiz around security providing agentic security across the stack. you'll see us
[10:43] front and center and being the cloud provider that is highly differentiated in this area. Uh that aside, this is what that st that looks like for you to be able to say okay is Google cloud the
[10:55] right platform for me to run AI workloads I would say yes and with that I pass it on to Muhit to actually walk you through the journey that InMob has had in this process.
[11:08] Thank you. Thank you Naveen. Uh so uh before I get to the slight little intro about myself uh I'm I'm group uh one of
[11:22] the founder and group CTO of Inobi. Uh Inobi today is basically it's uh includes three companies. So the first one is uh we are one of the largest uh marketing platform where you have supply
[11:34] side and demand side uh platform together. Then we have a consumer tech called glance. What we are trying to do is uh that we believe that discovery or search discovery is almost coming to an end and we will move into uh agent
[11:47] eiccommerce uh shopping era. So we are trying to redefine that genre over there. We are seeing some great results and then we are actually one of the largest weather app. Uh so we own one
[11:59] weather which is the second largest weather app after uh you know the weather channel and uh uh we have a very close partnership with weather channel as well. So these are the three very different business that come together but what is common uh for all these
[12:12] three businesses the underneath infrastructure layer that we have uh while they all three have different nuances uh little bit about our scale is so our user reach across the globe is around 1.6 6 billion uh it's quite a
[12:25] large number. We are present all the way from Japan to US. So every country in between every continent in between we have uh our business uh roughly 1.7 billion SDK reach through various
[12:37] publisher and user and as you can see we have around 320 billion ad queries 780 million impression every day 170 countries served and almost 11 big exchange integrated including Amazon uh
[12:50] Google DB3 uh and then we have we are presence around 22 countries so it's a fairly large um uh you know operation that we have as a company we are also very
[13:02] hands-on we actually ran our own data center all the way till pandemic hit us around 2019 or uh 20th and the main reason was if you see we serve user from US to Japan and everything in between
[13:15] and all these market are at a very different price point so if I have to give you a example of let's say commerce shopping on glance if I have if I'm spending $10 on AI to generate your
[13:27] image your styling and all it is is completely okay because you know the ARPU in US and developed world could be as high as $50 as soon as you go to India it is actually below 10 cents so nothing works the AI as we know it today
[13:41] in that shape and format doesn't work at that price point so we have to do lot of engineering in between and I'll share some of the work that we have been able to do with data bricks and uh Google uh to make that happen but this is the
[13:54] scale that I'm talking about almost uh tens of terabyte data ingested every hour you know uh we have data center or rather region GCP region almost in every continent and they constantly talk to
[14:07] each other because if you are running a campaign and you want to spend $10 the $10 can spend anywhere you know and how do you know that you are not spending $1,000 in a minute where you lose your shirt so the the the serving will stop
[14:20] as soon as you finish your $10 and how it is done is basically you know the information exchange all across multipetabyte of data process for intelligence
[14:32] algorithm machine learning and whatnot. Roughly thousands concurrent job 10,000 peak spark nodes that we have and uh roughly hundreds of monthly platform user internally who are doing some uh algorithm and machine learning work over
[14:45] here. So how do we how do we prioritize our own life? You know when we think of engineering is nothing but the manifestation of our own strategy. So I try to put a very simple four uh bucket
[14:59] where we see things. One is look AI is everywhere today we are not asking the question that uh you know where to put AI in fact the question that we are asking with every team why you are not
[15:11] using it and it doesn't matter it's not just engineering in fact you'll be surprised that most of my enabler function uh all the way from legal finance to accounting they are completely uh digitalized and AI ready
[15:24] today. uh the first time I lost my uh the kota of token per user was not the engineering it was actually finance team uh so so that's the first part you know
[15:36] whether it's GTM pre-sale everything we want to make sure that you know AI is doing all the heavy lifting uh as you can see we have lot of businesses you know from glance one weather and SSP and DSP uh the biggest problem that a
[15:50] organization like us have is the information is silos people don't talk to to each other and you can't optimize humans but I can optimize the information exchange and I can make it sort of seamless that you know whenever
[16:02] you are looking at something you are getting a holistic view rather than looking into your business so this is second part is fairly important to us and we spend lot of time making sure this is happening intelligence should be a commodity across group uh that's
[16:15] that's what we do over here and third one is cost I think cost is a very important factor with us as I was telling you you know we we have been doing this on our on our own and we learned a way of running things at a
[16:28] unit economy which is actually at best in my industry. So give you a sense there is one metric that every engineer keeps in his head is I by R which is my infrastructure uh cost per revenue and no matter what
[16:41] you write it can never go beyond 3%. So 3% you can ask where it come up. We just you know we we we we come up with the equation that we feel comfortable my CFO and as a CTO uh but no code is being
[16:55] written at Inobi where you will exceed your infrastructure to revenue cost by 3% and next year we will try to take it to 2%. So if you have those kind of those kind of uh uh constraint I think
[17:07] your engineering practices changes quite a lot you know and that is some of the work that you know we do uh in this uh category and it's quite important. Fourth is basically really important we deal with lot of information and
[17:20] advertising itself is a highly regulated industry you know we have to deal with GDPR CCPA in California DPDP in India so every country has its own region its own laws and you can't handle them manually. So it needs to be coded, it needs to be
[17:33] into the system. So trust and compliance and governance was a very big part of it and I think we struggle for quite some time until we latch on to databics where it comes as part of the uh offering.
[17:49] So uh what do we have in AI in production? So we have you know so at the same point of time we have lot of structure but I also want to make sure that my team have autonomy to do things because AI things are changing so fast. I I don't want this to be a top-driven
[18:03] strategy and philosophy. In fact, every engineer should be able to take a call that hey, this makes sense. Let me try it out and if it works, then they can share it with the other team as well. So, what you're seeing on the point over here is we have lot of fragmented AI
[18:15] experimentation. Uh there is a bpoke pipeline model as well. You know, something is in production that requires certain level of stabilility. So you go into that area as well and loosely connected but at the same point of time
[18:28] highly integrated cataloging across the business so that you know there is no rigid connection over there but it you will be able to make sense of all the information when um you know uh you you are looking at the information. So that
[18:41] that that was one of the one of the simplification that we brought as a as a challenge early on. Uh crossbu intelligence as I told you system is never a problem. is how your org structure get defined basically gets
[18:53] replicated in your engineering system and I think we were no different so as we were launching more and more business this became a very big problem with us uh we just didn't know what is happening in next uh corner you know because just
[19:05] they sit in a different team uh cost discipline as I was telling you you know has been a very important factor for us uh so how much idle time is there in my uh uh VMs where we are heavily loaded
[19:18] where we are not uh BCP compliant and whatn not. So all that uh you know challenges come over here. We were actually at a point of running uh I don't know if it's mentioned over here five or six version of data warehouse.
[19:30] So it's like you you're running pipelines again and again and you just filling these data warehouse all over the game from Vertica to snowflakes to everything was there uh if you look back at around 2020 in the company u trust and compliance was a very loosely uh
[19:43] concept you know in people's head not coded as a system. So I think these are the four big challenges uh you know uh that that when we uh saw the system and started our hyperscaler journey um these
[19:57] were the four uh key challenges that we had. So why Google cloud? So I think we have been testing hyperscaler since 2014. I think 2015 Azure came to us and they
[20:10] said hey why you run your own data center and all those things but none of their system were ready. I think we had crashed even the PC stage. So I think it took some time for the hyperscaler to be at a level where they are as reliable as
[20:23] they they are and and and the progress that they have made in last five six years is just unbelievable and one of the things which was very important for Google is uh when the pandemic happened
[20:35] it was very clear in our head that we don't want to be in this business because it's taking huge amount of time at least one of the metric that I was looking at is where is my 80% time or mind share is going and it cannot be
[20:47] procuring server and setting them up and running running them you know that means I'm spending 20% uh time on my business so that was very clear but one of the biggest issue was who gives me the flexibility to meet the fourth
[21:00] constraint that I was talking to about you know without doing least amount of work cost I still have the flexibility and I'm not losing out you know what is happening out there in the world so I think Google was
[21:12] uh one offering that we felt give me enough kind of robustness and scale that I needed and at the same point of time it was flexible enough for us to basically run things the way we want to
[21:24] run it. So GKE and spot VMs I mean Google Google team comes to us and they give us hard time because we are so uh efficient in hogging up all the spot that they have
[21:38] at some point of time 90% of my infrastructure could be running on VMs and I don't know why I share that screen with Google they keep on increasing the price of VMs. Uh but but jokes apart this is the kind
[21:50] of flexibility I I I get with Google. Very flexible economy you know I can work on every component and agree on a price point and work with them work with Google you know and they they they partner us not only just to basically uh
[22:03] you know provide services as a vendor but to optimize it because you know they they see the value in that because then they can uh give it to other folks as well. Uh Vortex and Gemini is our way to go if you want to launch something faster in the market. We always use
[22:16] these two product and we launch it over there. Uh storage as I told you there is a huge amount of information that we generate it passes uh across the globe. So I think we don't need to do anything. We are completely on GCS and everything
[22:28] is taken care. um manage data services, data procs, serverless, spark I think the usual suspect that you know everyone could be using enterprise security and network I think this is also a very
[22:40] important component and especially appstack and all is getting very important for us and now viz and everything is becoming part of Google I think it became very seamless for us so I think all the six four six seven things that we used to worry about I get
[22:53] uh right at the back almost 60% of our bandwidth had been freed where we don't have to worry about any of those things. Uh then once again coming back to data bricks as I was telling you um 50% of uh
[23:07] inobi engineers are actually ex Yahoo. So we actually inherited lot of Hadoop expertise. We were writing Hadoop we were writing yan we were writing spark we were doing everything. So before data
[23:20] bricks was actually founded we knew mati and Ellie everyone because we working very closely with them. you know when the spark was still in Berkeley labs uh two or three of my my classmates and junior they have worked at data bricks.
[23:32] So there was a natural uh uh working relationship with this company before databicks even became what it is today. But at the same point of time you know what database provide to us is unified
[23:45] lakehouse. There is a open format I can use anything uh you know iceber and it just provides all the support out of the back. uh as once again elasticity is given uh cost awareness as I tell you
[23:58] never leaves our life. So datab bricks also falls into the same thing. Unity catalog and governance. This is the amazing feature that data bricks has. Uh honestly, it took us one year to actually realize the capability of this component and it was a great boon. You
[24:11] know, we really have to worry about nothing from lineage to governance to everything. It just comes uh out of the box. Orchestration pipeline. You know, my data data science team, they love it. Earlier, you know, we were running our
[24:23] own MLOps uh and and they have to sit with two three engineers to basically take their model to production. Sometimes they would wait three weeks. And to me that was a criminal waste of time. Now they can take their own model you know they can train it test it
[24:36] offline take it to production uh in matter of days. Um basically end to end ML life cycle you know you can offload it to datab bricks. The last but not least I think the data bricks and GCP itself they are great partners. So you
[24:49] know I think the way I see it I don't make a distinction that I'm talking to two different team. most of the time they are together understanding our environment understanding our problem and then uh working towards that so this is really u you know how the things uh
[25:03] look for us when I say 3% infrastructure I by R it gets divided to these two and we in the middle uh so do data bricks autoscaling also happen on spot almost 30 to 70% of my jobs even to this day
[25:17] run on it uh and we realize maximum amount of saving over there uh all the new features are available to us through Vert.Ex and Gemini inference all all the table we don't have to do any kind of engineering on top of it and as I said
[25:31] any new uh uh open-source software that we are using whether it's Polaris iceberg it's it's all supported over there so it's a very well integrated uh system in a nutshell so we have basically multiple Gemini
[25:45] services in our exchange uh demand classification context contextual platform Jarvis is our own basically uh uh S sur or rather ops uh software that
[25:57] we have written you know that can tell us what is going on with the system and all. I want to spend a little bit more time on the supply categorization and commerce intelligence because commerce is a very big part of uh advertising and
[26:09] what is happening as I was telling you we are moving from discovery to agentic era. So in discovery and display advertising what used to happen earlier there used to be fixed asset you know here is the banner here is the video just go and serve it. So most of your
[26:21] elbows were basically regression or mathematical equation you know I will anticipate the CTR I will anticipate the CPD and I will show you the banner commerce the world has changed a
[26:33] merchant will come let's say Amazon will come to you and say hey here are the thousand SKU that I want to advertise today and here is the price point here is the rorowass you know nothing about those SKUs their ads are getting
[26:45] dynamically created their price is changing and their context is changing who they want to target So every template, every modeling that is happening, it's happening on the runtime. So if you are seeing a video, chances are that video was estu
[27:05] is not in the advertising campaign anymore. Something new is over there. So how do you do that? I think for that I guess you know we use Gemini to basically stitch and do all those kind of creation over there. Same thing supply categorization. You know our ads goes into various sites, various app.
[27:18] It's not enough to say hey this site belongs to gaming or this site belongs to uh betting or news. You need to understand what page it is what is happening over there and most of the time these assessments are happening at
[27:30] a runtime. So our SDK right now carry a small LM kind of component as well who can actually create a context and who can tell you at the runtime that what is happening. So lot of these heavy lifting I think we offload to Gemini and it has
[27:43] it has been a incredible uh upside for us. Uh then there are some you know uh the last item is pretty interesting. It's Plutus when I say 3% I by R. We are
[27:55] actually quite serious about it. It calculates at a runtime that how systems are performing and what is the cost that is happening and we use lot of Google APIs lot of data bricks and Google APIs at a runtime and then give a holistic
[28:07] cost over there. So um this is a typical architecture AI application compute engine and cataloging governance and storage how there I think you uh I'm assuming you guys all know about it and
[28:19] um so this is how the workflow uh looks like the event lands in GCS ETL in EDI runs over there we have a tier system on that bronze system bronze gold and uh uh
[28:32] silver because you want to treat the data the the kind of seriousness it has and the kind of priority it So every system depends on what category they belong to you know because you could be losing cost doing a real-time processing for the data that doesn't
[28:45] change for a day versus not doing the real-time processing for the data they're changing so dynamically uh all the discovery metadata we use Apache uh iceberg data lake tables uh along with the data bricks as I said
[28:58] depending on what category you belong to we use vertex we use click house we use some in-house train tool as well to basically give you the reporting that you need And last but not least is recommendation risk and all the machine learning pipelines. So this is a sort of
[29:12] sort of assembly line that you see on the workflow that we have. Uh this is basically uh you know a a sort of a view on the outcome. Uh we have been able to basically improve our governance query
[29:25] to up to 80% just by putting this flow and structure together. Uh absolutely 20% reduction in the duplication. You know three jobs are creating same pipeline. Now you know the fraction layer will figure out this information
[29:37] is already ready and we don't need to run that job. This was pretty amazing. The data set onboarding and access approval and whatn not used to take days for us. Now it can happen in hours and I think we should be able to
[29:49] bring it under our I think within a month or so. I think this is a humu uh massive efficiency unlock for us. uh same thing you know from 10% to 90% data is right now governed inside the
[30:02] immobi which is a very big deal considering inobi is also planning to go through IPO journey and this was the one of the important component uh pre-sales I think my my business team loves what we have been able to do the kind of
[30:14] information that we have been able to provide them you know their productivity their uh you know output has gone as you can see 48 to 90x and these are real number you know they have they have benefited so massively around it um some
[30:28] account coverage and you know supply chain categorization Gemini as I was telling you 90% of our creative we are not involved in it they are get they gets created on its own and most of the time they're bang on you know they are exactly performing or they beat any
[30:41] manual campaign creation today but I think I think when we talk about AI I know the LLM and model they take unfair amount of mind share from us that
[30:56] doesn't take credit away from the hard work that goes in the engineering and putting this stack together. So I what I'm trying to highlight over there is using AI using Gemini is one part of it
[31:08] but I think when you look at the overall picture what it what to run it effectively to run it at a price point where it is meaningful to us meaningful to our you know customer I think there's a lot more going uh goes over there. So
[31:21] first one is obviously you know open formatted federated catalog uh you know cloud choices config decision not migration is fairly important to us uh whatever software we choose because we
[31:33] work with many other many as I was telling you we are present in almost every country we have a stack on Ali cloud as well so how do we want to write everything once again on that cloud or we want to write something that can work over there I think this philosophy comes
[31:46] very handy over there because if I'm running something on Apache and Polaris is completely portable with data bricks over here as well and I can run it as it is on Alleloud as well. Um Google cloud all elasticity you know I think Google
[31:58] has done amazing job governance everything is offloaded to data bricks within the cost and discipline and the price point that we want. Uh last but not least is uh the foundation the way
[32:10] the AI run in production is one part but is it auditable? Can I govern it? Can I manage it? And can I deliver the faster result which are way exponential beyond the manual uh baseline is very
[32:22] important. You know AI is not something uh that basically is designed to give you incremental change. You know that is what the engineering used to be uh doing 20 years from now. But AI I think it needs to change uh your life massively
[32:36] in a way you have never imagined. And I want to call out two particular example that I'm quite proud of. Uh so one of the first team that actually moved to AI using uh Gemini was legal team. So my
[32:48] legal team goes through a roughly 30,000 to 40,000 contract every week and they would read and then you know one contract could be thousands of pages. They would read every page red line and see you know where the liabilities can
[33:01] pass on to Inobi and what is written over there and you can throw even 200 lawyers and they would not be able to finish the job and we were at that point and that team came to us we build a system using Gemini completely now
[33:13] almost all the contract as soon as they come into the system the initial red nailing is done in one minute so by the time is reaching to that legal uh uh employee the contract is very well
[33:26] written read and highlighted everything where they need to pay attention or if something got skipped, something is not missing, they can actually go and tell the model, hey, these are the four points that were not captured and here are the changing dynamics and changing
[33:39] scenario. If DPDP or CCPA come with a new rule, we will just uplate we we will we will just update this system and it will start to read those contract accordingly. So I think this was one of the massive change. Today what used to
[33:53] take them almost month to finish 30,000 40,000 contract they can actually do it in few hours. So that is that is the number that you know I want to call it out over there. Um engineering has no
[34:05] choice you know so they have to basically uh join the AI bandwagon and doing over there. But I think I feel quite proud of non-engineering team and how they are uh doing over there. So that that is all what I have uh what we
[34:17] have been able to achieve with putting uh databics and uh Google together and uh thank you.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.